Received-SPF: pass (google.com: domain of linux-kernel+bounces-191817-linux.lists.archive=gmail.com@vger.kernel.org designates 2604:1380:40f1:3f00::1 as permitted sender) client-ip=2604:1380:40f1:3f00::1;
From: xu xin <xu.xin.sc@gmail.com>
To: xu.xin16@zte.com.cn
Cc: bigeasy@linutronix.de,
	john.ogness@linutronix.de,
	linux-kernel@vger.kernel.org,
	linux-rt-users@vger.kernel.org,
	liu.chun2@zte.com.cn,
	rostedt@goodmis.org,
	si.hao@zte.com.cn,
	yang.yang29@zte.com.cn,
	zhang.yunkai@zte.com.cn
Subject: Re: [PATCH  5.10-rt] printk: ignore that console preempted by irq/softirq
Date: Tue, 28 May 2024 06:40:03 +0000
Message-Id: <20240528064003.776618-1-xu.xin16@zte.com.cn>
In-Reply-To: <20240523235537910_yxfGMbTcSOmMkcfuK2d8@zte.com.cn>
References: <20240523235537910_yxfGMbTcSOmMkcfuK2d8@zte.com.cn>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

> > diff --git a/kernel/printk/printk.c b/kernel/printk/printk.c
> > index 7f27cfee283e..faab85dd4439 100644
> > --- a/kernel/printk/printk.c
> > +++ b/kernel/printk/printk.c
> > @@ -3735,6 +3735,14 @@ bool pr_flush(int timeout_ms, bool reset_on_progress)
> > diff = 0;
> > 
> > for_each_console(con) {
> > + /*
> > + * When we're in the unpreemptible context on the same cpu
> > + * with which the thread of console locates on, we should
> > + * ignore this console, because it's a vain.
> > + */
> > + if (!preemptible() && con->thread &&
> > + task_cpu(con->thread) == smp_processor_id())
> > + continue;
> > if (!(con->flags & CON_ENABLED))
> > continue;
> > printk_seq = read_console_seq(con);
> 
> This does not apply.
> There is `may_sleep' set earlier.
> 
> There is no console_lock() around for each…
> 

Sorry, I don't get it.

To clarify it again, this patch aims to solve the useless waiting of pr_flush
when the console is preempted by the current irq/softirq. This has nothing to
do with might_sleep().

> The other question is which kernel started enforcing might_sleep() for
> pr_flush(). This should be applied to all kernel or none so we don't
> have random behaviour across kernels (5.4 yes, 5.10 no, 5.15 yes).
>

Sorry, my understanding is that pr_flush didn't start enforcing might_sleep().
This patch can apply to 5.10 and 5.15 where the problem exist.

> This is a delay of max 1 sec during bug() and panic(). Not sure how
> "critical" this is…

In some industrial control scenarios, bugs and warnings containning a
pr_flush delay of 1 sec is very critical to the upper services.

Especiall for watchdog timeout(< 2s), just WARN can easily lead to system reset,
which is unacceptible.